Report

The Structure of Networks with emphasis on information and social networks T-214-SINE Summer 2011 Review Ýmir Vigfússon Part I Chapters 2, 3 and 5 Graph Theory Graph ◦ Nodes connected by edges Path ◦ Set of nodes where each consecutive pair is joined by an edge Component ◦ Set of nodes where each two nodes are connected by some path Distance ◦ Length of shortest path between two nodes Breadth-first search (BFS) ◦ Method to find shortest paths from one node to all others Breadth-first search (BFS) From the given node (root) ◦ Find all nodes that are directly connected These are labeled as “distance 1“ ◦ Find all nodes that are directly connected to nodes at distance 1 If these nodes are not at distance 1, we label them as “distance 2“ ◦ ... ◦ Find all nodes that are directly connected to nodes at distance j If these nodes are not already of distance at most j+1, we label them as „distance j+1“ Breadth-first search (BFS) Breadth-first search (BFS) Strong and Weak Ties Triadic closure ◦ If two people have friends in common, they‘re more likely to become friends themselves Clustering coefficient ◦ Over every pair of my friends, what fraction them are friends themselves? Local bridge ◦ If A and B are friends and have no other friends in common Strong and Weak Ties Suppose edges are now either strong (friends) or weak (acquaintances) Strong Triadic Closure Property (STCP) ◦ For every node A with strong ties to some friends B and C, then B and C are connected by an edge (either weak or strong). Thm: ◦ If a node A satisfies STCP and is involved in at least two strong ties, then any local bridge it is involved in must be a weak tie Strong and Weak Ties People seem to maintain a small number (~40) of strong ties (Facebook/Twitter) Structural hole: ◦ The „empty space“ or „missing“ edges between groups that don‘t interact much Betweenness of an edge: ◦ Essentially the number of shortest paths (between pairs of nodes) that use the edge But if there are multiple shortest paths between A and B, say k of them, the contribution to betweenness of intermediate edges is only 1/k instead of 1 Positive and Negative Relationships Suppose all edges exist, and are either + (friends) or – (enemies) (Strong) Structural Balance: ◦ For every set of three nodes, either three + edges or exactly one + edge. Balance Theorem: ◦ In a balanced complete graph, either all nodes are friends or it can be divided into two coalitions of mutual friends such that people in different groups are enemies Positive and Negative Relationships Weak Structural Balance: ◦ No set of three nodes have two + edges and one - edge. Weak Balance Theorem: ◦ In a weakly balanced complete graph, nodes can be divided into groups of mutual friends such that people from different groups are enemies Positive and Negative Relationships What if not all the edges exist? Thm: ◦ A signed graph is strongly balanced if and only if it contains no cycle with an odd number of negative edges Part II Chapters 6, 7, 8, 9 Game Theory Game theory deals with the outcome of person‘s decisions in various situations ◦ How to choose among different options ◦ People‘s choices when interacting with others Def: A game consists of three things. ◦ (1) A set of players ◦ (2) Each player has set of options how to behave called strategies ◦ (3) For each choice of strategies, each player receives a payoff from the game. Game Theory Your partner You Presentation Exam Presentation 90,90 86,92 Exam 92,86 88,88 Strategies Payoff for row player,column player Prisoner‘s dilemma ◦ Each player thinks studying for exam is safer, but this doesn‘t reach the social optimum Game Theory How are the players likely to behave? Assumptions ◦ The payoff for a player should capture all rewards Including altruism ◦ Players know all possible strategies and payoffs to other players ◦ Players are rational Each players wants to maximize her own payoff Each player succeeds in picking the optimal strategy Game Theory Best response ◦ The best choice of one player, given a belief of what the other player will do E.g. „U is a best response to L“ Dominant strategy ◦ If a strategy is the best response to every strategy of the other player ◦ We can assume it will be played! Game Theory Nash equilibrium ◦ For strategy S by player 1 and T by player 2, the pair (S,T) is a Nash equilibrium if S is a best response to T, and T is a best response to S ◦ More generally, a cell in the payoff matrix is a Nash equilibrium if no player wants to unilaterally (i.e. by himself) deviate to an alternative strategy A game might have multiple Nash equilibria, or none. Important Games Battle of the sexes ◦ Which kind of movie to rent? Romance Thriller Romance Thriller 1, 2 0, 0 0, 0 2, 1 ◦ Two equilibria, but which one will be played? Hard to predict the outcome ◦ Depends on social conventions Important Games Stag Hunt ◦ If hunters work together, they can catch a stag ◦ On their own they can each catch a hare ◦ If one hunter tries for a stag, he gets nothing Hunt Stag Hunt Hare Hunt Stag Hunt Hare 4, 4 3, 0 0, 3 3, 3 Two equilibria, but “riskier“ to hunt stag ◦ What if other player hunts hare? Get nothing ◦ Similar to prisoner‘s dilemma Must trust other person to get best outcome! Game Theory Mixed strategies ◦ Each player now commits to a strategy with a certain probability e.g. Player 1 has probability p of choosing strategy H Nash‘s Theorem: ◦ When we allow mixed strategies, Nash equilibrium always exists. Social optimum ◦ The cell in the payoff matrix that maximizes the sum of the players‘ payoffs. Mixed Strategies: Example Penalty-kick game ◦ Soccer penalties have been studied extensively Left Defend left Defend right 0.58, -0.58 0.95, -0.95 Right 0.93, -0.93 0.70, -0.70 ◦ Suppose goalie defends left with probability q ◦ Kicker indifferent when (0.58)(q) + (0.95) (1-q) = (0.93)(q) + (0.70) (1-q) ◦ Get q =0.42. Similarly p=0.39 ◦ True values from data? q=0.42 , p=0.40 !! The theory predicts reality very well Evolutionary Game Theory Game theory continues to apply even if no individual is overtly reasoning or making explicit decisions Natural selection ◦ More fit organisms will produce more offspring ◦ This causes genes that provide greater fitness to increase their representation in the population But how do we measure „fitness“? Evolutionary Game Theory Key insight ◦ Many behaviors involve the interaction of multiple organisms in a population ◦ The success of an organism depends on how its behavior interacts with that of others Can‘t measure fitness of an individual organism ◦ So fitness must be evaluated in the context of the full population in which it lives Analogous to game theory ◦ Organisms‘s genetically determined characteristics and behavior = Strategy ◦ Fitness = Payoff ◦ Payoff depends on strategies of organisms with which it interacts = Game matrix Evolutionary Game Theory What happens at equilibrium? ◦ Nash equilibrium doesn‘t make sense since no individual is changing their strategy Evolutionarily stable strategy ◦ A strategy is evolutionarily stable if everyone uses it, and any small group of invaders with a different strategy will die off over multiple generations Evolutionary Game Theory Organism 2 Organism 1 Theorem S T S T a, a c, b b, c d, d ◦ Strategy S in a two-player, two-strategy symmetric game is evolutionarily stable when either (i) a > c or (ii) a = c and b > d Theorem ◦ If strategy S is evolutionarily stable, then (S,S) is a Nash equilibrium. ◦ If (S,S) is a strict Nash equilibrium, then strategy S is evolutionarily stable. Routing Games Players (drivers) commuting from some node s to t All players are making private decisions about what path to drive to minimize latency ◦ Latency of an edge depends on the number of drivers Q: What happens at equilibrium? ◦ Don‘t care how we got into the equilibrium Routing Games Braess‘s Paradox ◦ A new road can hurt performance at equilibrium Routing Games Theorem ◦ The average latency of traffic at Nash equilibrium is at most 2x the latency of optimal traffic Assumes linear latencies on edges („ax+b“) A better bound of 1.33x is known. ◦ In other words: If people were somehow told what routes they should drive, the average latency would at most 33% better than when everyone selfishly picks a route Auctions Selling one item, but don‘t know how much it‘s worth Each bidder (player) has her own intrinsic value for the item. ◦ Willing to purchase it up to this price ◦ Values are independent Auctions help us discover this value Goal: Maximize social welfare ◦ Sell item to the person who wants it the most In terms of own value, not the bid! Auctions Dutch auctions ◦ Price gradually decreased until somebody offers, and gets the item at that price Sealed-bid 1st –price auctions ◦ Person who wrote the highest bid in an envelope gets it at that price Properties ◦ Dutch auctions are strategically equivalent to sealed-bid 1st –price auctions ◦ They are not truthful: underbidding is a dominant strategy for all players E.g. bidding (n-1)/n fraction of your value (in uniform distribution) when there are n players ◦ They maximize social welfare (i.e. efficiency) Auctions English auctions ◦ Price gradually raised until all bidders drop out, person with last offer given the item Sealed-bid 2nd –price auctions (Vickrey) ◦ Person with highest bid gets item, but pays second highest bid Properties ◦ English auctions are strategically equivalent to Vickrey auctions ◦ They are truthful: people bid their true value ◦ They maximize social welfare in equilibrium. Auctions How much does the seller make? 1st -price and Dutch auctions: ◦ Bidders reduce their bids by a factor of (n-1) /n ◦ Expect largest bid to be n / (n+1) Expect revenue: (n-1) / (n+1) 2nd price and English auctions: ◦ Seller commits to collecting less than max. bid ◦ Look at highest and second-highest bids Expect revenue: (n-1) / (n+1) Known as revenue equivalence Part IV (and V) Chapters 13 - 15, 18 Sponsored search Need to auction multiple ad slots with different properties If we knew buyers‘ valuations, we could use a matching market Sponsored search Theorem ◦ We can always set prices so that if buyers buy the item they most want, all items are sold (market clearing prices) Done by gradually raising prices until there is a unique perfect matching between buyers and sellers ◦ Market clearing prices always produce socially optimal outcome But we don‘t know the valuations... ◦ Want to encourage truthful bidding Sponsored search Google‘s Model: Generalized Second Price Auction ◦ Slot i is assigned to the ith highest bidder at a price per click equal to the (i+1)st highest bidder‘s bid ◦ But this does not encourage truthful bidding! However, it may give Google higher revenue even though this is not well understood Sponsored search VCG mechanism ◦ Idea: Each individual is charged the harm they caused the rest of the world Theorem ◦ Truthful bidding is a dominant strategy in VCG for each buyer ◦ The VCG assignment is socially optimal It maximizes the total valuation of any perfect matching of slots and advertisers Structure of the web Hyperlinks on the web form a directed graph ◦ In the early days, links were navigational ◦ Now many links are transactional E.g. when you click on your shopping basket Two types of directed graphs ◦ DAG (Directed Acyclic Graph) If u can reach v, then v cannot reach u ◦ Strongly connected Any node can reach another via a directed path Theorem ◦ Any directed graph can be expressed in terms of these two types Structure of the web Strongly connected component (SCC) ◦ Every pair of nodes can reach each other ◦ No larger set with this property Theorem ◦ Every directed graph is a DAG on its SCCs On the web, there is just one giant SCC Ranking web pages How should we sort search results? Observation ◦ Should do link analysis, i.e. consider links to be votes Hubs and authorities (HITS) ◦ Quality as an expert (hub) Total sum of votes of pages pointed to ◦ Quality of content (authority) Total sum of votes of experts ◦ Principle of repeated improvement ◦ Theorem HITS converges to a single stable point ◦ Sort results by decreasing authority score Ranking web pages Hubs and authorities depend on the topic (e.g. „cars“). Expensive to compute. PageRank ◦ A page is important if it is pointed to by other important pages Again, principle of repeated improvement ◦ On what websites will a random web surfer spend his time? ◦ PageRank(i) = probability the random web surfer visits page i at any given time Sort result pages by decreasing PageRank ◦ Fact: PageRank is the first eigenvector of the adjacency matrix PageRank example Ranking web pages Properties of PageRank ◦ Problem: Spider traps and dead ends Places where you can‘t get back to the SCC Solution: Do random jumps (teleports) to any page about 15% of the time ◦ Ranking is independent of the topic unlike HITS Could do personalized PageRank by teleporting to topic-relevant page instead Power laws It is not unlikely to find someone who has double your wealth (or salary) ◦ Far less likely to find someone who‘s double your height (i.e. normal distribution) Such distributions are heavy-tailed ◦ E.g. Amazon‘s sales rank Power laws Look at the in-degree distribution of the web ◦ Very many pages have a small degree ◦ But reasonbly many pages with an insane number of in-links Power-laws ◦ Bulk of the degree distribution shows up as a line on log-log scale ◦ The fraction of pages that have k in-links is approximately 1 / k2, a power-law ◦ More generally: if the fraction of (population) that has k (something) follows 1 / kα then power-law with exponent α Power laws Exponent typically between 2 and 3 for any popularity measure (many examples!) Power laws Why do power laws arise? Preferential attachment model ◦ A.k.a. rich-get-richer model ◦ Idea: People have a tendency to copy decisions of people who act before them ◦ Web pages arrive sequentially With probability p they: Link to a previous page uniformly at random With probability 1-p they: Link to a page with probability proportional to that page‘s indegree ◦ Theorem: This model produces a network with a power-law degree distribution The exponent is 1+1/(1-p) Part VI Chapters 16,17, 19 and 20 Information cascades Herding or information cascade ◦ You join a crowd even if it contradicts your private information ◦ Could be rational: Others might have more information than you do! Cascades may not lead to optimal outcome ◦ First few people could be wrong Cascades can be based on very little information ◦ People start ignoring their private information once the cascade begins Cascades are not robust ◦ Only a bit more information can break it Emperor‘s new clothes Collective action Some decisions have direct benefit effects ◦ The „network effect“ – buying a fax machine is better for all previous fax machine owners Model where everybody sees all actions ◦ Each person i has a threhold ti ◦ Will adopt behavior if at least ti other people are adopters ◦ F(x) = fraction of people with threshold ≤ ti Can simulate the dynamics to find a fixed point ◦ When F(x) = x, no more changes to adoptions Network effects Number of adoptions will increase to the next fixed point F(x) x (number of adopters) Number of adoptions will decrease to the next fixed point Network effects But model ignores the network How should we organize a revolt? ◦ Personal threshold k: I will show up if at least k people show up. ◦ People know the threshold of their friends Each person is playing a 2-person game with each of his friends ◦ Great if both choose same option ◦ Payoff: a if both choose A, b if both choose B ◦ Fact: Node will choose A if b/(a+b) fraction of friends choose A Network effects Theorem ◦ The use of a new technology A spreads monotonically (i.e. nobody will switch back) Small world phenomenon People are connected by short paths ◦ Milgram experiment: six degrees of separation ◦ MSN experiment: average of 6.7 hops Surprising because social networks tend to have high clustering coefficients ◦ This impedes the exponential growth (i.e. assuming my 100 friends each have a 100 distinct friends, etc.) Small world models Suppose all edges are created independently with probability p ◦ The Erdos-Renyi random graph model Properties Small world models Theorem ◦ The diameter of a random 3-regular graph on n nodes is O(log n) with high probability. ◦ Gives us a working definition of what „short paths“ really means: O(log n) But Erdos-Renyi random graphs don‘t capture triadic closure as in social networks ◦ Modify model to have lots of local edges, and a few random links Small world models Watts-Strogatz model ◦ Goal: Create a random graph with lots of triangles (local edges) and small diameter ◦ Start with a regular lattice (e.g. ring or a grid) ◦ Rewire: For each edge with prob. p, move the other end to a random vertex ◦ Theorem Diameter is O(log n) Small world networks Small world networks Observation: People are able to find these short paths in a decentralized way! ◦ Geographic navigation should work in polylogarithmic time Geographic navigation: Forward the message to the node I know closest to the destination ◦ Not captured by the Watts-Strogatz model Theorem: Takes at least n2/3 steps to reach target Need a model such that geographic navigation algorithm works naturally Small world networks Kleinberg‘s model ◦ The random edges in the Watts-Strogatz model were too random for navigation Create long-range edges with probability ◦ P[u links to v] = 1 / distance(u,v)α distance(u,v) is the grid distance between u and v ◦ High α: Mostly know people close by ◦ Low α: Know people from very far away Theorem ◦ For α=2, geographic navigation takes O(log2 n) time ◦ For α≠2, no algorithm can navigate in poly-log time Actually it‘s α=dimension of the grid Small world networks Kleinberg‘s model fits real-world data well ◦ Using ZIP-codes in LiveJournal and measuring distance by rank (# people living closer), get an excellent fit to 1 / rank(u,v) Implications for peer-to-peer networks ◦ Can construct networks that allow for efficient search for content Chord from MIT is the best known ◦ Efficient: find a file using O(log n) messages